[libc-commits] [libc] [llvm] [libc] Extend fputil::sqrt to use floating point instructions for arm32. (PR #134499)

[via libc-commits]

lntue wrote:

> The AMD docs say that f32 sqrt has 1ULP and f64 has (2**29)ULP accuracy. Currently AMDGPU uses the builtins, so we could probably remove `src/math/amdgpu` and same for nvptx.

are `__builtin_elementwise_sqrt` and `__builtin_sqrt` the same on AMDGPU and NVPTX?

https://github.com/llvm/llvm-project/pull/134499